Where does the bug appear (feature/product)?
Cursor IDE
Describe the Bug
In one long-running Agent chat, the parent coordinator is in Multitask Mode. When the user says “Start multitasking”, the parent launches a background Task with:
• run_in_background: true
• resume: “self”
• prompt: You are the forked subagent; continue executing your task.
Those forked workers often never produce an assistant turn. The subagent transcript gets only the injected user messages and then freezes. The UI still looks like work is in progress. After ~100 minutes, Cursor logs rpc.run error=true for that composer. A later retry can handshake and even start a tool call, then stall again. No files are edited.
A non-fork background worker (full task prompt, subagent_type: generalPurpose, no resume) can start: it writes one assistant message and issues tools (UpdateCurrentStep, GetDynamicTools), then freezes. The jsonl is not updated again. Request traces show processPartialToolCall with no completed rpc.run span. Target source files stay untouched.
This is not an SMB/EBUSY filesystem stall and not a hung product test process. Status-check workers in the same chat complete normally.
Steps to Reproduce
- Open Cursor Agent chat in Multitask Mode on a local project. Model: Grok 4.6.
- Give the parent a multi-step implementation task (or let it already be working).
- Send: Start multitasking
- Observe the parent call Task with resume: “self”, run_in_background: true, and prompt You are the forked subagent; continue executing your task.
- Wait. Open the child composer / inspect agent-transcripts//subagents/.jsonl.
- Optional variant: after a hung fork, have the parent spawn a new background Task with a full prompt (no resume). That child may emit one assistant+tool turn, then freeze the same way.
Expected Behavior
• The forked background subagent should run a first model turn immediately, use tools, and either finish or keep writing the transcript.
• The parent / UI should not present a silent, never-started worker as in-progress work.
• If rpc.run fails, the subagent should surface an error and stop, not sit hung for hours.
Operating System
MacOS
Version Information
Cursor 3.17.21 stable
Commit 8f2a112cb2845a97b75fd932ea5c470579ca4060 (realCommit …4063)
product.json date 2026-08-25T01:05:08.089Z
VS Code 1.128.0
OS macOS 26.3 (Build 25D125), Darwin 25.3.0, Apple Silicon arm64 (T6041)
App /Applications/Cursor.app
Model grok-4.6 (Cursor Grok 4.6), maxMode=false, idSource=selectedModels[0]
Mode Agent + Multitask Mode
Privacy unknown to this report author — please note Privacy Mode vs Share Data when posting
How version was obtained (do not invent): cursor --version printed 3.17.21 / 8f2a112cb2845a97b75fd932ea5c470579ca4060 / arm64; /Applications/Cursor.app/Contents/Resources/app/product.json matched.
For AI issues: add Request ID with privacy disabled
Subagent (composerId) Request ID Notes
59380492-83aa-4a09-894a-6db437dc8855 75be26bb-a589-4bae-b25c-9793c61c4171 Fork hang; rpc.run error=true twice
1df5a1b5-4060-4ec0-937f-a58e1cde86d2 96429f02-1abe-4165-b2b9-9094d14b9756 Fork hang; rpc.run error=true ~109 min
26bd25e6-d5ed-4736-8837-01b5ed883ffa 1c0f2240-f864-4a41-a5e8-a4b1ace9377f Started tools then silent; three rpc.run error=true
52b2a686-0ae4-4ceb-afb2-eda4f6e9b3d4 cd9f19bc-7a42-4c09-8bf9-ebbe3bcf5230 Non-fork restart; first tools then freeze; no completed rpc.run span as of 08:50 PT
Parent conversation / composer: 6cf40309-8e6e-494c-a9c0-f6694ac2a206
Additional Information
Evidence table (local transcripts)
Paths under ~/.cursor/projects//agent-transcripts//subagents/. Home anonymized.
Agent ID Title (parent Task description) Spawn (PT) Last jsonl mtime (PT) Size / lines Assistant after kickoff? Outcome
59380492-83aa-4a09-894a-6db437dc8855 Fix pdfcompare vector attributes 06:12:54 06:13:44 1709 B / 3 lines No Hung ~2h; confirmed by status worker at 08:10
1df5a1b5-4060-4ec0-937f-a58e1cde86d2 Build html2pdf 1.0.24 (2nd fork) 06:07:10 06:07:56 581 B / 3 lines No Never started
26bd25e6-d5ed-4736-8837-01b5ed883ffa Build html2pdf 1.0.24 (1st fork) 05:46:54 05:47:43 1873 B / 4 lines Yes: first Reads/Grep only Transcript frozen; packages later appeared on disk anyway
52b2a686-0ae4-4ceb-afb2-eda4f6e9b3d4 Fix pdfcompare vector attrs (restart, full prompt) 08:15:23 08:15:29 3696 B / 2 lines Yes: 1 turn (UpdateCurrentStep + GetDynamicTools) Still frozen at 08:50+ PT (~35 min); target C files unchanged
Contrast (same parent, completed):
Agent ID Title Notes
6facaca6-fbbd-4c58-8bda-d918fe3c78b3 Check pdfcompare worker status Full prompt, background; completed (turn_ended success)
a32448c7-7b99-4e59-a3fb-56eb8dbefd29 Check 1.0.24 build status Completed
fa362dbb-51b9-461b-b4e6-09c412b22b49 Kill hung test-suite terminal Completed (unrelated CEF/html2pdf test hang)
bac530bd-b56a-4852-b6d6-f854e4695c08 html2pdf width tiling Earlier non-fork worker; did complete (41 lines)
Exact kickoff text that hung forks received
Three consecutive user messages, same timestamp as spawn:
- The original user task (implementation prompt).
- Start multitasking
- You are the forked subagent; continue executing your task.
Parent Task call shape (from parent jsonl):
{
“description”: “Fix pdfcompare vector attributes”,
“prompt”: “You are the forked subagent; continue executing your task.”,
“resume”: “self”,
“run_in_background”: true
}
Same shape for “Build html2pdf 1.0.24”. Parent once noted “Self-fork failed” and fell back to a full-prompt background Task (that fallback historically worked).
Renderer log (model bind)
2026-08-28 05:46:54 [info] [buildRequestedModel] composerId=26bd25e6-… catalogModelId=grok-4.6
2026-08-28 06:07:10 [info] [buildRequestedModel] composerId=1df5a1b5-… catalogModelId=grok-4.6
2026-08-28 06:12:54 [info] [buildRequestedModel] composerId=59380492-… catalogModelId=grok-4.6
2026-08-28 08:15:23 [info] [buildRequestedModel] composerId=52b2a686-… catalogModelId=grok-4.6
Request-trace excerpts (rpc.run error=true)
From ~/Library/Application Support/Cursor/logs/20260827T100810/window1_wb2/output_20260827T113925/cursor.requestTraces.log:
59380492 — spawn 13:12:54Z; first rpc.run failed after ~102 min
2026-08-28T14:55:26.934Z rpc.run composerId=59380492-… requestId=75be26bb-… error=true durationMs=6152153
2026-08-28T15:17:18.607Z rpc.run composerId=59380492-… requestId=75be26bb-… error=true durationMs=1308564
1df5a1b5 — spawn ~13:07Z; rpc.run failed after ~109 min
2026-08-28T14:55:56.198Z rpc.run composerId=1df5a1b5-… requestId=96429f02-… error=true durationMs=6525664
26bd25e6 — multiple rpc.run failures on the same requestId
2026-08-28T13:00:08.508Z rpc.run composerId=26bd25e6-… requestId=1c0f2240-… error=true durationMs=794247
2026-08-28T14:55:55.549Z rpc.run composerId=26bd25e6-… requestId=1c0f2240-… error=true durationMs=6943721
2026-08-28T15:16:46.868Z rpc.run composerId=26bd25e6-… requestId=1c0f2240-… error=true durationMs=1244627
52b2a686 — last activity ~15:17:27Z processPartialToolCall; NO rpc.run completion by 15:50Z
After some rpc.run failures the client retried (handshakeEstablishment, writeInitialRequest, even toolCallStarted) without writing a successful assistant completion to the jsonl.
Hypothesis
- Primary: Multitask self-fork (resume: “self” + injected “continue executing” / “Start multitasking”) often never delivers a first model turn to the child. The composer is created (buildRequestedModel succeeds) but the agent RPC (rpc.run) sits until it errors (~100 min). Transcript remains user-only.
- Secondary: Even a fresh background Task can stall after the first tool batch (GetDynamicTools / MCP move_agent_to_root, or Reads). Traces show processPartialToolCall without a finished rpc.run. Jsonl is not updated; work does not continue.
- Not confirmed: permission-prompt disappearance (a known Multitask bug) — these hangs happened with no shell yet, so likely not waiting on Allow for a terminal.
- Not the cause: SMB EBUSY on network paths; hung html2pdf/CEF run-tests.sh (separate, killed in terminal 211764).
Impact
• Work silently never starts. User waits tens of minutes to hours.
• Parent reports the child as launched / in progress.
• Implementation (here: vector attribute comparison in a C tool) never happens; source mtimes unchanged.
• Status-check subagents work, which makes the hang look like “the worker is busy” rather than “the worker is dead.”
What this is NOT
• Not a hang inside the user’s C compiler or test suite (no pdfcompare compile/test processes; no new terminals).
• Not SMB/EBUSY on /Users/…/Network/… (no file writes were attempted by the hung pdfcompare workers).
• Not the 12-hour html2pdf CEF test-suite hang (different pid/terminal; already killed).
Related / similar reports (not proven identical)
• The subagent just hung on “thinking”
• Sub Agents wont run and are stuck at “New Subagent - Starting up” — partly a UI “Starting up” bug (said fixed in 3.2.11); this report is transcript + rpc.run stall on 3.17.21
• Background subagent can get stuck when permission prompt disappears — related Multitask stuck state; our hung forks never reached a shell Allow prompt
• Early compression of the 1M Subagent context window (broken subagent resume) — resume-related, different symptom
This specific combo (Multitask + resume: “self” + “Start multitasking” / “forked subagent; continue executing” + rpc.run error=true + empty jsonl) was not found as a documented, fixed issue on 3.17.21.
Does this stop you from using Cursor
Sometimes - I can sometimes use Cursor